Open-Weight AI Models Match Proprietary Systems in Network Hacking Tests

Open-weight artificial intelligence models are closing the capability gap with tightly controlled proprietary systems in executing complex network cyberattacks, according to new research published in November 2026 [1]. The findings arrive from experiments conducted by AI measurement company Vals, which tested various models on multi-step instructions across mock corporate networks [1].

Testing Network Penetration Capabilities

Vals evaluated different AI models to measure how effectively they could carry out complex, multi-step network instructions across simulated corporate environments [1]. The testing framework involved two distinct scenarios [1]. The first scenario required models to target a mock corporate network accessible via a public-facing website [1]. The second scenario involved breaching two separate company networks sequentially, requiring the model to compromise the initial network to gain entry to the second [1].

According to the research, OpenAI's most advanced cyber model, Daybreak—which remains restricted to trusted partners—achieved the highest performance by completing 81 percent of assigned tasks across an average of two tests [1]. However, Chinese open-weight model DeepSeek closely followed, successfully completing 73 percent of objectives on average [1]. Researchers noted that both models achieved identical scores on several specific tasks [1].

Cost Efficiency and Open-Weight Advantages

The study highlighted significant cost disparities between operating proprietary models and utilizing open-weight alternatives [1]. Running an attack using GPT-6-Sol incurred token costs ranging from $17 to $46, whereas executing equivalent operations via DeepSeek cost between $2 and $3 [1]. Vals CEO Rayan Krishnan noted that this substantial cost difference creates a strong financial incentive for defensive practitioners to adopt open models over proprietary ones [1], given that open models can also be operated on local infrastructure [1].

Industry Implications and Defense Concerns

The rapid maturation of open-weight hacking capabilities challenges previous assumptions held by major US frontier artificial intelligence laboratories [1]. Companies like Anthropic had premised restricted release strategies—exemplified by Claude Mythos Preview introduced in April—on the assumption that gated access controls would successfully buy time before advanced hacking tools became widely accessible [1]. Vals researcher Nikil Ravi stated that open models that anyone can download are now beginning to match the best frontier models locked within corporate labs [1].

Despite these technological milestones, real-world observations of state-backed and criminal hacking activity indicate that malicious actors frequently rely on older, less capable open and proprietary models to conserve resources [1]. Alex Zenla, cofounder and chief technology officer of cloud security firm Edera, noted that while open-weight models can generate advanced vulnerabilities and attacks—particularly when safety training is disabled by modifying weights—closed labs likely possess unreleased models that exceed current public capabilities [1].

Major frontier AI labs have historically restricted access to their most advanced models to prevent the proliferation of sophisticated cyber weapons. Beginning with Anthropic's Claude Mythos Preview in April, developers utilized gated access programs to buy time for defenders, while projects like Anthropic's Project Glasswing aimed to provide security teams with a head start. Concurrently, cybersecurity throughout the year has seen AI automate core tasks to help both attackers and defenders discover code vulnerabilities and probe network weaknesses.

The erosion of the capability gap between open-weight and closed AI models signals a major turning point for the cybersecurity landscape in India and globally. As advanced network-penetration tools become inexpensive and broadly accessible to the public, organizations must reevaluate their defensive postures against automated cyber threats, while security professionals can leverage cheaper open-weight models for defensive testing.

The widespread availability of advanced hacking capabilities in open-weight models may accelerate the frequency and sophistication of automated cyberattacks. Defenders may increasingly adopt cheaper open-weight architectures for security testing, while the traditional moat of proprietary AI labs regarding cyber utility may permanently diminish.

It remains unconfirmed whether AI has fundamentally altered the conceptual foundation of cybersecurity or simply amplified existing threats. Additionally, closed labs may possess unreleased models with superior capabilities that are not reflected in current public tests.